WorldFoundry
Open-source infrastructure for world models, unifying inference and benchmark evaluation across video generation, 3D/4D representation and embodied-action models behind one TUI, CLI and Studio interface.
WorldFoundry is open-source infrastructure for world models: a shared stack for staging local model assets, running inference through a TUI, CLI or a Studio interface, and evaluating results against benchmark layouts across video generation, 3D/4D representation, embodied action and interactive worlds.
Its inference path integrates a range of third-party world models and vision-language-action policies, with a unified attention-backend selection across FlashAttention, SageAttention, xFormers and SDPA, Triton kernel compilation and caching, and multi-GPU context and sequence parallelism. The Studio interface handles model discovery, Conda environment isolation, GPU allocation and distributed launch, and presents the results.
The recommended workflow is to bootstrap the environment and stage checkpoints and datasets, generate and inspect artifacts, then score against a benchmark once the artifacts match its expected layout.